Machine Unlearning on WMDP chem forget-set
56BF16 ScoreZero-shot
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Zero-shotBackbone=Qwen-3-8B2026.05 | 56 | — | — | |
| Zero-shotBackbone=Llama-3.1-8B-Instruct2026.05 | 53.3 | — | — | |
| LUNARBackbone=Qwen-3-8B2026.05 | 52.1 | 53.4 | 0.013 | |
| Global GABackbone=Llama-3.1-8B-Instruct2026.05 | 49.3 | 47.3 | -0.02 | |
| LUNARBackbone=Llama-3.1-8B-Instruct2026.05 | 48.1 | 49.7 | 0.016 | |
| Surgical GABackbone=Llama-3.1-8B-Instruct2026.05 | 42.7 | 42.3 | -0.003 | |
| MANSUBackbone=Llama-3.1-8B-Instruct2026.05 | 33.3 | 30.7 | -0.027 | |
| Surgical GABackbone=Qwen-3-8B2026.05 | 31.3 | 31.7 | 0.003 | |
| MANSUBackbone=Qwen-3-8B2026.05 | 30.7 | 27.4 | -0.033 | |
| GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 27.3 | 27.3 | 0 | |
| NPOBackbone=Llama-3.1-8B-Instruct2026.05 | 25.3 | 22.7 | -0.027 | |
| SimNPOBackbone=Qwen-3-8B2026.05 | 24 | 27.7 | 0.037 | |
| Global GABackbone=Qwen-3-8B2026.05 | 23.7 | 29.3 | 0.057 | |
| NPOBackbone=Qwen-3-8B2026.05 | 23.7 | 23.7 | 0 | |
| SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 23.3 | 23.3 | 0 | |
| GU+SimNPOBackbone=Qwen-3-8B2026.05 | 23.3 | 23.3 | 0 |